Apple Foundation Models 3 (AFM 3)

What it is

Apple Foundation Models 3, usually written AFM 3, is the third generation of the models Apple uses for Apple Intelligence. Apple published it on 8 June 2026 ^[1]. The previous generation had two models. AFM 3 is a family of five, which lets Apple choose a model for each device and each kind of task.

The family has two groups: models that run on the device and models that run in Apple's cloud. The on-device models run on the phone or computer itself, and the server models run on Apple's servers, where the Private Cloud Compute system handles requests that are too heavy for a device ^[1].

Origins

The models were built in collaboration with Google. Apple says they were custom-built with Google and pre-trained on the latest generation of cloud TPUs, Google's accelerator chips ^[1]. This collaboration followed an agreement that Apple and Google announced in January 2026, under which Gemini would underpin Apple's next-generation models ^[4].

When Apple launched AFM 3, Craig Federighi said the new models do not use Gemini's models or infrastructure. Wikipedia, citing MacRumors, reports that the models were trained with reinforcement learning and refined using outputs from Gemini frontier models ^[3][2]. Readers should keep these two accounts separate. Apple's statement describes the models themselves, while the second account describes the training process as reported by the press.

Key features

AFM 3 Core is an on-device model with 3 billion parameters, using a dense design in which every parameter takes part in each computation ^[1].

AFM 3 Core Advanced is a natively multimodal on-device model. It is a sparse model of 20 billion parameters, and each request activates between 1 and 4 billion of them. It keeps the full model in flash storage, which is the phone's long-term storage, and loads the routed experts into memory only when they are needed. Apple uses a technique it calls Instruction-Following Pruning to make this possible ^[1]. Sparse design means that only part of the network runs for a given input, so a large model can be used on a device with limited memory.

On the server side there are three models. AFM 3 Cloud is an upgraded version of the Parallel-Track mixture-of-experts design, running on Private Cloud Compute, Apple's system of servers for requests too heavy for a device. ADM 3 Cloud (Image) handles image generation and editing. AFM 3 Cloud Pro is the most capable of the group and is aimed at agentic tool use and complex reasoning. Agentic tool use means the model plans a series of steps and calls tools to carry them out ^[1].

Apple also worked with Google and NVIDIA to extend Private Cloud Compute to NVIDIA GPUs running in Google Cloud. AFM 3 Cloud Pro is optimised for these GPUs ^[1]. A GPU, or graphics processing unit, is the kind of chip that most AI systems now use for training and running models.

The models handle several kinds of input and output: text, image understanding, image generation and editing, and audio, which covers text-to-speech and dictation ^[1].

Apple reported its own evaluations. Human graders preferred AFM 3 Cloud over the 2025 server model on 64.7% of prompts, against 8.7% for the older model. They preferred AFM 3 Core over the 2025 on-device model on 45.6% of prompts, against 23.3% ^[1]. These are Apple's own evaluations, so they show how Apple measured its models and not how outside testers would rank them.

Some of the advanced on-device features need the newest hardware. Apple's materials point to an iPhone 17 Pro, or a Mac with an M3 chip and 12 GB of memory ^[1][2].

Where it sits

AFM 3 powers Siri AI, which Apple announced on the same day for iOS 27 ^[2]. It follows the second generation, described in the 2025 entry. Because the models are recent, the material about them is still limited, and this entry will need updating as Apple publishes further detail.